Neural-Bayesian Structure Learning for Discrete Choice Modeling

Conventional discrete choice models typically treat traveler attributes as parallel, independent inputs, failing to account for directed dependencies (e.g., how income influences vehicle ownership). This paper introduces Neural-Bayesian Structure Learning (Neural-BSL), a framework that couples differentiable structure learning with random-utility-based choice estimation. By using a particle-based variational inference approach, the model learns a directed acyclic graph (DAG) representing dependencies among traveler attributes while simultaneously estimating the choice model. The observed choice is maintained as an alternative-specific utility comparison, preventing the categorical outcome from distorting the recovered attribute structure.

Neural-BSL offers three primary advancements. First, it provides an end-to-end differentiable procedure where the discovered attribute structure and the choice model inform each other. Second, it introduces a utility specification where individual-level coefficients are weighted by the learned DAG, allowing the structural context to influence choice probabilities. Third, it enables structural policy simulations: when an attribute is intervened upon, the model propagates the change through the learned DAG in topological order, updating downstream attributes before recomputing choice probabilities. This allows researchers to observe not just the change in mode share, but also the downstream traveler or trip adjustments that drive those responses.

Evaluated on stated-preference data from Seoul and revealed-preference data from London, Neural-BSL achieves predictive performance comparable to multinomial logit and neural network benchmarks. Beyond prediction, the model recovers behaviorally coherent dependency structures, such as the link between driver's license possession and car ownership. In policy simulations, Neural-BSL reveals that interventions often produce different substitution patterns than static models. For instance, in the Seoul dataset, license surrender among older travelers leads to a shift toward subway transit, a response that static models fail to capture because they do not account for the cascading loss of private vehicle access.